Build and own a Total Loss Specialist Model.
Is this damaged vehicle or property a total loss, and what should the settlement be?
Overview
Speeds up total-loss determination by comparing repair estimate to actual cash value automatically instead of manual adjuster review.
Who owns this decision
VP Claims, Total Loss Manager
Why owned AI may fit
Total-loss formulas and thresholds vary by state/jurisdiction and must be applied precisely and consistently.
The verdict
A total-loss engine freezes claim evidence at a point in time, runs external systems to assess damage, compute repair costs, value the vehicle, and price salvage. Oumi extracts facts from estimates and comparables, explains what the external outputs mean, finds conflicts, compares repair versus total-loss scenarios, drafts claimant-facing language, and expresses confidence and abstention. Qualified adjusters remain accountable for determination, offer, and payment. Photos aren't complete inspections. Repair costs aren't settlement truth. Valuations aren't settled. Your auditors and claimants must trust the process, not a model.
Before and after
Most adjusters reconcile evidence scattered across photos, repairs, valuations, and salvage bids. A supplement arrives later, or comparables age. An owned total-loss engine freezes the evidence snapshot, runs specialist systems (vision, estimating, valuation, salvage), applies rules deterministically, and uses Oumi to extract and explain what those outputs mean. Qualified adjusters review the packet without recreating it.
Frozen evidence prevents conflicts, slippage, and bias
When repair, valuation, comparable, and rule data arrive over different times, mismatches go unnoticed. An estimate says $45K repair; valuation says the vehicle is worth $40K; a week later a new estimate says $48K. Freezing claim state at a point in time forces resolution of these conflicts before analysis starts. Oumi detects when VINs, mileage, condition, or prior damage don't align. If identity or jurisdiction is unresolved, it abstains.
Sompo's published case describes production AI checks in an image-based total-loss workflow. CCC documents adoption of automated valuations. These confirm feasibility. They don't prove settlement fairness, jurisdiction compliance, or reduced dispute rates. Your own audit trail of frozen packets, external outputs, reconciled facts, adjuster decisions, and claimant responses is the only proof that matters.
What you own, what stays rented
Buy
- Claim administration systems
- Policy and coverage determination
- Computer vision and inspection
- Repair estimating and parts pricing
- Market valuation and comparable data
- Salvage pricing and auctions
Own
- Time-correct evidence graph
- Frozen packet contracts
- Fact extraction and conflict detection
- Scenario comparison and explanation
- Adjuster feedback and corrections
- Dispute and appeal resolution memory
The owned asset is the reviewed claim-to-vehicle-to-damage-to-estimate-to-comparable-to-valuation-to-salvage-to-rule-to-determination graph. A photo score, accepted offer, or auction result alone isn't the asset. You retain the right to audit, override, correct, appeal, litigate, and learn from every step.
Prove it before you bet on it
Freeze 5,000 to 50,000 historical claims across repair, total-loss, and borderline cases split by jurisdiction, vehicle class, and damage type. Validate that identity, policy, photo, estimate, valuation, comparable, salvage, rule, and outcome data are complete and time-correct. Benchmark current manual operations, external vision, estimating, valuation, and salvage separately. Run read-only shadow comparison with Qwen3.5-9B and NuExtract 3 on identical frozen packets. Measure fact accuracy, scenario coverage, citation fidelity, and abstention. Only when adjuster burden stays flat and severe errors drop should you permit assist mode.
Model candidates to evaluate
- ModernBERT (Apache 2.0)
- Qwen3-VL (Apache 2.0)
- Qwen3.5-9B (Apache 2.0)
These are candidates, not performance claims. Run them against your evaluation set.


